Next Generation Private Equity : The Prospect Edition

How to use Machine Learning — in the right place and at the right time

Despite Machine Learning (ML) generating more interest than ever, there are times when the risk of deploying it ineffectively can actually outweigh its benefits, leading to failed investments and frustrated goals. By stark contrast, businesses that plan ahead and incorporate ML at the right place, and at the right time, can enjoy outsized returns. As part of WovenLight’s ‘Next Generation Private Equity’ series of exploratory articles, our Chief Scientist Aris Valtazanos takes a look at the current landscape, and shares concrete recommendations from experience.

Author: Aris Valtazanos, Chief Scientist at WovenLight.

Machine Learning (ML) is a powerful tool that can unlock value across modern businesses. Recent leaps in computational power and data storage capacity have significantly accelerated the speed and scale at which algorithms and automation can drive better business outcomes, and moreover, the proliferation of modern data platforms has made ML more accessible to industries that were not previously amenable to data-driven disruption [1][2].

But with great power comes great responsibility. As with other powerful tools, the risks of failing to utilise ML effectively can actually outweigh its benefits, leading to failed investments in tools and technology, unsuccessful projects, and broader organisational dissatisfaction. There is no one-size-fits-all approach to deploying ML.

At WovenLight, we advise businesses to plan ahead and incorporate ML at the right place and time. So how do we make crucial implementation decisions around ML?

The first step is to break down the characteristics of the problem. We work with a set of questions to ensure ML plays the right part in the ultimate solution, split into three categories: problem and hypotheses, data, and solution.

1. Problem and hypotheses

Is the business problem clearly defined?

Before diving into the details of what a solution, the precise business problem must be identified. If the problem is formulated too broadly (e.g. “increase customers”), it can lead to inappropriate solutions being developed downstream. Instead, focusing on a specific business-driven KPI or operational lever (e.g. “reduce rate of customer churn”) helps narrow down the solution space and improves alignment.

What are the existing solutions to this problem inside the business?

If a problem is perceived as important, the business may already have an operational solution in place — possibly without analytics or automation. For instance, in the churn reduction problem above, the sales team might be prioritising customers based on a simple heuristic (e.g. value of products owned)…or even randomly.

Mapping the existing solution space helps us to understand the status of the KPI that a potential ML solution would seek to improve; it also ensures that there is strong buy-in around any new ML-driven initiative, and that it doesn’t risk crossing wires with existing efforts.

What are the hypotheses around this business problem?

Another key step to inform the implementation of an ML solution is to formulate a list of business hypotheses around it. A business hypothesis like this is a statement that asserts the impact of an attribute or factor on the desired business outcome: for example, “customers with high consumption are more likely to churn”.

A large and diverse set of hypotheses is likely to benefit from a ML-driven solution, which can uncover the most significant ones and therefore point to the optimal business actions downstream.

Who are the end users of the resulting solution?

The people developing an analytics solution are not always those who will end up using it. Remembering this is key to identifying the intended audience ahead of time, engaging with them in design, and incorporating their feedback.

2. Data

What data exists to cover these hypotheses?

Modelling the hypotheses depends on the availability of data — internally in the business, and/or from external and public sources. If some hypotheses are not captured by the available data, the resulting models may be less insightful.

Is data coverage sufficient?

In order to be useful, the available data needs to provide good coverage across time (e.g. multiple years to cover different seasonal patterns) and entities of interest (e.g. completeness across customer base). If either of these conditions is not met, the feasibility of the solution, or the size of the opportunity, will be impacted.

Is the data quality sufficient?

Data should also adhere to certain quality criteria in order to be practically usable. There are several openly available tools (see [3]) that address common quality metrics.

Another key dimension is identifying which metrics impact your ability to deliver a good solution to the problem. While a lot of data sets appear to be of poor quality at first glance, parts of them can be still useful, or there may be data transformations that can raise the quality to an acceptable level.

Additionally, the rigour of the data collection process itself impacts the quality of the data itself. For example, if key sales metrics are recorded only by a fraction of the sales team (or on ad-hoc basis), data will be sparse and incomplete.

What other data initiatives are underway in the business?

Many organisations are undertaking other data integration initiatives while exploring the use of analytics and ML; for example, bringing all their data together into a data lake. So we need to consider the impact these data integrations could have on the analytics efforts, and the opportunities for mutually beneficial synergies. There is a common perception that ML-based projects can only be engaged once all data is integrated and cleaned, but this is not always true.

First, data integration initiatives are often ongoing or delayed, so also postponing ML-driven value creation for these reasons can lead to loss of value. Second, the scope of data integration initiatives is typically broader than anything analytics and ML could cover (e.g. supporting Business Intelligence work). Third, and most importantly, data integration initiatives can be accelerated by analytics projects. For example, certain parts of data cleaning and merging can be undertaken as part of the analytics project, and then fed back to the wider data integration effort.

Alternatively, analytics projects can point to high-priority data sets that can add value across the business, and thus aid prioritisation within the data integration initiative itself.

3. Solution

What type of (analytical) solution will create the most value for the business?

Once there is a good understanding of the problem and the current (benchmark) solution to it, the next step is to determine whether ML can improve upon it and to clarify which solution would create most value for the business.

These can be split into four broad categories, in rough order of complexity of implementation:

• A descriptive approach (which doesn’t involve ML) focuses on presenting data and insights to experts in a more structured way, e.g. through dashboards. This is recommended when humans are already perceived as good decision makers, but could benefit from access to data and insights in a more structured way. Descriptive outputs can also be suitable when decisions depend on human expertise that does not get captured in the data, or is difficult to replicate through automation.
• An explanatory model takes a collection of analytically defined hypotheses and uses ML to identify which are the most significant — which in turn can be translated to business actions. This approach is useful when the end goal is the transformation of a business process (e.g. improving how sales teams are run based on identified bottlenecks) and where it is important to quantify the relative importance of different factors (e.g. experience of sales people vs workload fragmentation).
• A predictive model forecasts what is going to happen and when. It helps provide personalisation, speed and automation; particularly in situations where volume and complexity is too high for humans alone to make fast and effective decisions (e.g. predicting which customers will churn).
• A prescriptive model takes the solution one step further by providing specific recommendations on *how* a desired outcome can be achieved. It can thus provide even more targeted recommendations; for example, finding the optimal amount of discount to offer a customer to prevent from churning.

These approaches need not be mutually exclusive; in fact, a successful solution may often involve a combination of them.

How complex does the analytical solution need to be?

Even when a business takes a predictive or optimisation modelling route, it needs to establish if there are simpler solutions that work well before jumping to a more complex ML model. A good starting step is to implement some heuristic approaches; for example, if the goal is to forecast sales over the next year, take last year’s sales (or a five-year average) as a heuristic solution. A simpler model can be more easily interpreted by end-users and is therefore more likely to be broadly adopted within an organisation.

How can an ML solution be integrated into an existing operational process?

Many good ML models fail to realise their potential because they’re not properly adopted on the operational side. In most cases, they are only one part of a larger puzzle, which includes business users, technology and tools, and other operational considerations. Therefore, it’s critical to identify the consumers of the solution and how the outputs can be structured to be useful to them.

How much ongoing maintenance and development will the ML solution require, and what resources are available to support this?

The life of an ML solution does not end when its initial development is complete. Efforts like integrating the model into live data systems so it can run on a continuous basis, and ongoing maintenance and development of new features. A business should ensure that it has the people with the right skillsets.

It’s indisputable that ML has become a powerful tool for many organisations, helping them uncover new value. But it’s also true that there are pitfalls for those companies that fail to implement ML effectively. By thinking thoroughly through all the facets of the problem, businesses can better prepare for success and ensure that the resulting solutions can create value in a sustainable manner.

WovenLight can help organisations on this journey

At WovenLight we have over ten years of experience helping organisations embark on their data and analytics journey. We are a multi-disciplinary team that brings together technical experts and investment professionals, who work closely together to identify and execute on opportunities for value creation through analytics. Through our exposure to multiple industries, we have created a playbook that can be adapted to the needs of a business, based on the dimensions identified in this article.

If you would like to find out more about what we do at WovenLight and how we can work with you, please reach out to us at info@wovenlight.com